is there a DJL best practice？

[iDonal]

because of torch-scatter and torch-sprase
DJL（https://github.com/deepjavalibrary/djl） cant use pre-build pytorch-lib
we should build pytorch_geometric and jni manually

[rusty1s]

Can you share some more information on this? I do not have any experience with djl yet.

[iDonal]

Djl can load multiple popular DL models： like MXNET TENSOR_FLOW PYTORCH using java language
For pytorch， the calling sequence likes ： java->jni->torch c++ front head
I successfully load torch_geometric model by torch c++ front head , but djl failed

[iDonal]

Exception:
Exception in thread "main" java.lang.RuntimeException: ai.djl.engine.EngineException:
Unknown builtin op: torch_sparse::ptr2ind.
Could not find any similar ops to torch_sparse::ptr2ind. This op may not exist or may not be currently supported in TorchScript.

[iDonal]

DJL official pytorch example: https://github.com/deepjavalibrary/djl/blob/master/jupyter/pytorch/load_your_own_pytorch_bert.ipynb

[rusty1s]

Have you seen https://github.com/rusty1s/pytorch_scatter#c-api? Any chance one can register a custom C++ library into djl? Other-wise I am not sure how this should be fixable :(

[iDonal]

Thanks! I try to learn something about JNI first. C++ example works.